Category: News

Student: Roberto Gomes de Lima

Title: Sistema de Localização Baseado em Pontos de Referência para Robô Pulverizador Autônomo em Estufas Agrícolas

Advisors: Glauco Fiorott Amorim and Cláudio Miceli de Farias

Committee: Glauco Fiorott Amorim (Cefet/RJ), Cláudio Miceli de Farias (UFRJ), Felipe da Rocha Henriques (Cefet/RJ), Pablo Rangel (IPqM), Diego Leonel Cadette Dutra (UFRJ).

Day/Hour: October 13, 2026 / 6 p.m.

Room: https://teams.microsoft.com/meet/2776857574829769?p=YUzBXudNh89g4i7Sus

Alternative room: https://meet.google.com/ovk-ducc-jgt

Abstract: The automation process in agriculture has been increasingly disseminated within the context of so-called Agriculture 4.0 (also known as Smart Farms), encompassing artificial intelligence, robotic manipulation, intelligent sensing, and autonomous navigation. The integration of these technologies has led to significant advances in recent years, enabling more efficient and precise agricultural operations. Enclosed environments with limited illumination, such as greenhouses, represent a particularly suitable scenario for automation, as they allow controlled conditions that support the deployment of autonomous systems and assist farmers in improving crop productivity. In this context, this paper proposes an autonomous navigation system designed for greenhouse environments, with the objective of guiding a mobile robot during spraying tasks. To enable reliable navigation in such GNSS-denied environments, the proposed system incorporates alternative positioning strategies based on short-range wireless sensing technologies, which provide absolute spatial references to complement proprioceptive sensors typically used in mobile robotics. The proposed method integrates UWB-based range measurements with onboard odometry and IMU data to estimate the robot’s position relative to fixed checkpoints. The architecture enhances localization accuracy and navigation stability while remaining less complex and more cost-effective than state-of-the-art approaches. Experimental validation shows a positioning error below 5.7 cm and a checkpoint error below 8 cm.

Student: Felipe Mello Fonseca

Title: Vulnerability Detection in Smart Contracts: A Study Using Automated Tools

Advisors: Diogo Silveira Mendonça and Pedro Henrique Gonzalez

Committee: Diogo Silveira Mendonça (Cefet/RJ), Pedro Henrique Gonzalez (UFRJ), Kele Teixeira Belloze (Cefet/RJ) e José Augusto Miranda Nacif (UFV)

Day/Hours: August 21, 2026 / 8:30 a.m.

Room: https://teams.microsoft.com/meet/290006386263688?p=ICWKLlepYzGWKoe2Ly

Abstract: Blockchain is a technology with applications across multiple domains, including finance, electronic voting, and the Internet of Things. In this context, transactions are frequently mediated by smart contracts — self-executing programs written primarily in Solidity. Because they handle significant cryptocurrency values, these contracts are frequent targets of  attacks, making automated vulnerability detection a relevant technical challenge. This work presents two complementary sets of experiments. In the first, we evaluate the evolution of two widely used static analysis tools Mythril and Slither applied to the SmartBugs Curated dataset under the DASP taxonomy, comparing results against manually validated ground-truth labels. In the second, we assess the performance of Gemini 2.0 Flash on the same vulnerability detection task, comparing it to results previously reported for GPT-4 in the literature. The results show that the latest versions of traditional tools represent concrete progress, particularly in reducing false positives, though limitations remain in specific vulnerability categories. LLMs demonstrated potential as auxiliary mechanisms in preliminary triage, especially given their accessibility, but still require human validation for auditing critical systems. Overall, the experiments reinforce that combining multiple approaches is the most promising path toward increasing the reliability of smart contracts prior to deployment on the blockchain.

Student: Felipe de Ávila Tavares

Title: Efeito da Cardinalidade e da Associação entre Atributos na Imputação de Dados Categóricos

Advisor: Jorge de Abreu Soares

Committee: Jorge de Abreu Soares (Cefet/RJ), Diego Nunes Brandão (Cefet/RJ), Ronaldo Ribeiro Goldschmidt (IME)

Day/Hours: August 25, 2026 / 2:30 p.m.

Room: Pavilhão I, sala P1-202 (laboratório 6)

Student: André Igor Pereira

Title: AranduFlow: Towards Using Performance and Domain Metadata to Predict workflows execution time in HPC Environments

Advisors: Rafaelli Coutinho (Advisor) and Daniel de Oliveira (Co-advisor)

Committee: Rafaelli Coutinho (Cefet/RJ), Daniel de Oliveira (IC/UFF),  Eduardo Ogasawara (Cefet/RJ), Victor Stroele de Andrade Menezes (UFJF)

Day/hour:  August 11, 2026 / 2 p.m

Room: https://teams.microsoft.com/meet/218267218561286?p=zCgPEPHZ496sanbuRF

Abstract: The increasing complexity of scientific experiments has expanded the use of workflows to organize applications executed in High-Performance Computing (HPC) environments. In these environments, predicting execution time in advance is challenging because of variations in input data, application parameters, and computing infrastructure configurations. Furthermore, resource contention in shared environments can cause similar executions to have different durations. To assist users with workflow submission and experiment planning, this dissertation presents AranduFlow, an adaptable science gateway that uses historical data and machine learning models to estimate execution-time ranges. The proposed approach integrates submission, metadata collection, and prediction mechanisms by combining input-data characteristics, domain-specific parameters, and computing infrastructure information. Although the architecture can be adapted to different applications, the predictive models are trained specifically for each workflow. The approach was evaluated using two real-world scientific workflows from the bioinformatics and astronomy domains, with distinct resource-utilization profiles, executed on heterogeneous HPC infrastructures. The experiments showed that tree-based models outperformed the reference method and achieved accuracies above 80% in scenarios using broader execution-time ranges. The results also indicated the relevance of domain metadata and workload characteristics to the predictions. When applied to a new computing infrastructure, calibration using a small set of local executions produced a substantial improvement in predictive performance. These results provide evidence that \texttt{AranduFlow} can support execution-time estimation for heterogeneous workflows and can be adapted to different computing environments.

Student: Edson Landim de Almeida

Title: Estimation of Soil Organic Carbon Using Quantile Approaches and Explainable Artificial Intelligence

Advisors: Diego Brandão (advisor) and Jorge Soares (co-advisor)

Committee Diego Brandão (Cefet/RJ), Jorge Soares (Cefet/RJ), Kele Belloze (Cefet/RJ) and Marcos Ceddia (UFRRJ)

Day/hour: May 15, 2026/ 9 a.m.

Room: https://teams.microsoft.com/meet/272684253086940?p=V1Gq8L6jg3r6sPZIuq

Abstract: Soil organic carbon (SOC) plays a central role in biogeochemical cycles, soil fertility, and climate change mitigation, constituting one of the largest carbon reservoirs in the terrestrial system. Although direct quantification of SOC using laboratory methods is accurate, it is costly and time-consuming and limits sampling density in large-scale surveys. In this context, statistical and machine learning techniques have been used to estimate carbon content from more easily obtained edaphic and environmental attributes. This study investigates the estimation of soil organic carbon using pedological data provided by the Brazilian Agricultural Research Corporation (Embrapa), integrating multiple linear regression, Random Forest, Quantile Regression Forests (QRF), and Explainable Artificial Intelligence techniques based on SHAP. Multiple linear regression was used as an interpretable baseline, while Random Forest was employed to capture nonlinear relationships and interactions among soil attributes. The quantile-based random forest approach enabled the analysis of different regions of the conditional distribution of SOC, considering the 0.1, 0.5, and 0.9 quantiles. Results showed that SOC estimation is characterized by high variability, asymmetry, and heteroscedasticity, with relevant differences among the evaluated land uses. The conventional Random Forest model captured general trends in the data but exhibited greater error dispersion in the higher-carbon ranges. The QRF model expanded the analysis by enabling the representation of lower and upper bounds on the response, providing a more informative interpretation of the uncertainty associated with the predictions. The SHAP analysis indicated the relevance of attributes related to soil color, depth, and texture, in agreement with pedological knowledge. Overall, the results indicate that integrating tree-based models, quantile approaches, and explainability techniques constitutes a promising strategy for estimating and interpreting soil organic carbon, thereby advancing methodologies for digital soil mapping and environmental monitoring.

Student: Ana Gabriela Viana de Araújo

Title: Application of concept drift-based methods for predicting goals in professional football

Advisor: Jorge de Abreu Soares

Committee: Jorge  de Abreu Soares (Cefet/RJ), Glauco Fiorott Amorim (Cefet/RJ), Pedro Henrique González Silva (UFRJ), Carlos Eduardo Ribeiro de Mello (Unirio)

Day/Hour: May 13, 2026 / 9 p.m.

Room: https://teams.microsoft.com/meet/235075217075845?p=ezgpJ8dNRKnMa8dTfM

Student: Daiane de Ascenção Cardoso

Title: Method for Assessing Accessibility in Biomedical Ontologies  Visualizations

Advisors: Kele Teixeira Belloze (advisor) and Felipe da Rocha Henriques (co-advisor)

BCommittee: Kele Teixeira Belloze (Cefet/RJ), Felipe da Rocha Henriques (Cefet/RJ), Ingrid Monteiro (UFC), Glauco Amorim (Cefet/RJ) and Luis Carlos dos Santos Coutinho Retondaro (Cefet/RJ)

Day/Hour: May 11, 2026 / 1:30 p.m.

Room: https://teams.microsoft.com/meet/218701639098481?p=ZeSHhNfSekBnHMS8F8

Abstract: Biomedical ontologies play a fundamental role in structuring and communicating scientific knowledge, yet representing them accessibly remains a challenge. This dissertation proposes a systematic and automatable method for evaluating the accessibility of ontology visualizations. A qualitative analysis through methodological triangulation (usability, communicability, and accessibility) revealed the absence of objective criteria in existing tools. Subsequently, 3,048 parameterized visualizations of the Hemoglobin class from the Sickle Cell Disease Ontology (SCDO) were generated under three contrast configurations. A sample was manually labeled using the PCHRD model (Perceivable, Comprehensible, Hierarchically clear, Reliably labeled, and Distinguishable), adapted from the Chartability framework, revealing a relevant combinatorial asymmetry: there are more ways for a visualization to be inaccessible than accessible. A Random Forest was applied in two iterations — a probe of accessibility prevalence (confirming the labeling finding) and a proof of concept with mPCHRD-driven sample curation —, indicating that visual attributes carry discriminative signal. To verify real-world applicability, a corpus of images was built from articles in the Journal of Biomedical Semantics, classified by multimodal models (Claude Haiku 4.5 and GPT-4o-mini) and manually evaluated. Finally, the Chain-of-Thought technique was adopted to inspect the models’ reasoning when applying the PCHRD criteria, comparing automatic classifications with human judgments. The results indicate that the method predicts accessibility levels from graphical features and that explicit reasoning makes the models more restrictive in identifying perceptual barriers, potentially supporting inclusive design practices in biomedical ontology visualization. This work offers a replicable protocol, formalizes PCHRD as an instrument for assessment and retrieval, and fills a gap in the quantitative measurement of accessibility in information visualization.

Student: Jorge Nelson de Souza Pavão

Title: Large Language Models for Detecting Collusion in Bidding Processes

Advisors: Kele Teixeira Belloze (Advisor) and Diego Nunes Brandão (Co-advisor)

Committee: Kele Teixeira Belloze (CEFET/RJ), Diego Nunes Brandão (CEFET/RJ),  Raissa dos Santos Barcellos (UERJ), Eduardo Bezerra da Silva (CEFET/RJ)

Day/Hour:  May 06, 2026 /  4:30 p.m

Room: Sala E-520 (Bloco E, 5º andar, Cefet/Maracanã)

Abstract: The practice of collusion in public procurement, characterized by illicit agreements between competing firms, generates significant economic losses, reduces the efficiency of public contracting, and undermines trust in institutions. Several studies have employed machine learning algorithms and statistical variables to identify indications of this type of fraud; however, challenges such as the scarcity of labeled data, difficulties in generalization, and inconsistent performance across different datasets still persist. In this context, this dissertation investigates the potential of Large Language Models (LLMs) for detecting indications of collusion based on numerical data from procurement processes. Models from different providers were evaluated using two strategies: prompt engineering and supervised fine-tuning. The experiments were conducted on three real-world datasets — Brazil, the United States, and Switzerland — and the results were compared with those of traditional algorithms reported in the literature. The findings show that prompt engineering did not produce satisfactory performance. With fine-tuning, the LLMs outperformed traditional algorithms in only one of the three evaluated scenarios. Considering performance, cost, and implementation complexity jointly, it is concluded that, at the present time, traditional algorithms offer a better cost-benefit ratio for the evaluated task, while LLMs may serve as a complementary alternative in specific scenarios where classical methods fail to achieve satisfactory results or in situations involving textual data.

Student: Balthazar da Silva Cunha Paixão

Title: Analisando a Robustez de Redes de Passes em Futebol usando Redes Complexas

Advisor: Glauco Fiorott Amorim

Committee:  Glauco Fiorott Amorim (PPCIC-Cefet/RJ),  Pedro Henrique Gonzalez Silva (PPCIC-Cefet/RJ) and  Claudio Miceli de Farias (COPPE/UFRJ)

Day/Hour: May 04, 2026 / 7 a.m.

Room: https://teams.microsoft.com/meet/258717362035218?p=8XGFKontHejSUpL3Ou

Alternative link: https://meet.google.com/icf-miyw-pzi

Abstract: This work investigates the relationship between the structure of passing networks in football and the competitive success of teams over a season. Starting from the hypothesis that structural properties of these networks are associated with competitive success, operationalized by final league standing, football is modeled as a complex system of collective interactions. Passing networks are represented as directed and weighted graphs, in which nodes correspond to players and edges represent passes exchanged between them. The IHG metric is proposed, grounded in the interquartile range of structural importance measures, designed to quantify the heterogeneity in players’ participation within the passing network. The empirical analysis considers five major European leagues — La Liga, Premier League, Ligue~1, Serie~A, and Bundesliga — in the 2015/16 season. For each team and match, passing networks are constructed in their baseline state, from which classical structural metrics are computed, followed by the application of progressive perturbations through node and edge removal until a minimum structural threshold is reached. The data suggest that IHG constitutes a stable indicator for associating network organization with competitive success. The results indicate that multiple topological properties are associated with teams’ final standings, although with different patterns across leagues. In particular, metrics related to global connectivity, structural cohesion, and reciprocity show stronger associations in leagues such as Serie~A and Bundesliga. In contrast, the Premier League exhibits lower sensitivity to traditional global metrics, with IHG emerging as the measure most consistently associated with performance, a result interpreted in light of the atypical nature of the 2015/16 season. The analysis under perturbation reveals that structural robustness is not uniformly distributed across metrics, as only a subset — particularly those associated with structural cohesion and, to a lesser extent, transport efficiency — preserves discriminative capacity after network degradation. The findings support the research hypothesis, indicating that competitive success does not depend on a single structural property, but rather on a set of topological characteristics whose relevance varies according to the competitive context. Finally, this work establishes an integrated framework for structural and robustness analysis in passing networks, enabling the evaluation not only of teams’ collective organization but also of their ability to preserve functional properties under perturbations.

Student: Vanessa Santos Soares

Title: Avaliação de modelos de aprendizado de máquina para a correção automática de redações segundo as competências do ENEM

Advisors: Eduardo Bezerra da Silva (advisor) and Gustavo Paiva Guedes e Silva (co-advisor)

Committee: Eduardo Bezerra da Silva (Cefet/RJ), Gustavo Paiva Guedes e Silva (Cefet/RJ), Geraldo Bonorino Xexéo (UFRJ) and Diego Moreira de Araújo Carvalho (Cefet/RJ)

Day/Hour: May 11, 20266 / 9 a.m.

Room: https://teams.microsoft.com/meet/23323520592245?p=fnhxrL9byDTK3YLLbi

Abstract: With the growth of remote education and the implementation of large-scale exams such as ENEM, the automation of essay grading has become an increasing necessity. This work investigates different machine learning strategies for the automatic evaluation of essays written in Portuguese, based on the five assessment competencies defined by ENEM. A total of 9,599 essays were analyzed, collected from the Vestibular Brasil Escola portal, covering 102 topics published between 2009 and 2024. Two main approaches are compared: (i) traditional methods based on TF-IDF and linguistically engineered features extracted from the texts, and (ii) pre-trained language models with fine-tuning (XLM-RoBERTa with LoRA). Model performance is evaluated using the Quadratic Weighted Kappa (QWK) metric, which measures agreement with human raters. The study aims to demonstrate that pre-trained models provide significant improvements in robustness and reliability, outperforming feature-engineering-based approaches. This research contributes to the advancement of Automatic Essay Scoring (AES) in Portuguese by offering a benchmark and comparative analysis that can support future studies and educational applications.

Student: Gustavo Melo

Title: Reconhecimento de Entidades Nomeadas em Relatos Criminais Informais com Apoio de Metadados Estruturados

Andvisors: Eduardo Bezerra da Silva and Karla Figueiredo

Committee: Eduardo Bezerra da Silva (Cefet/RJ), Karla Figueiredo (UERJ), Gustavo Paiva Guedes (Cefet/RJ), Kele Teixeira Belloze (Cefet/RJ) and Ronaldo Ribeiro Goldschmidt (IME/RJ)

Day/Hour: May 12, 2026 / 9 a.m

Room: https://teams.microsoft.com/l/meetup-join/19%3ameeting_OWQ3ZTg5YzEtZjRmZC00NjkwLTg0MWUtYTBhODdkYTQwYTA0%40thread.v2/0?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%22049f2b6a-7ad4-4096-a3ed-db846537c488%22%7d

Abstract: This work investigates the problem of named entity recognition in informal crime reports recorded by the Disque Denúncia service. These reports, often marked by colloquial language, spelling mistakes, and free text structure, pose significant challenges to the use of traditional Natural Language Processing models. In addition to free text, the reports are accompanied by structured metadata, such as type of occurrence, location, and date which can provide additional relevant context for the task. In this study, we propose an approach based on the fine-tuning of large language models, using a manually annotated corpus with entities of the types Person, Location, and Organization. To overcome the scarcity of labeled data, the methodology includes the application of pseudo-labeling on a second, significantly larger corpus, thus expanding the training base in a semi-supervised manner. Additionally, the metadata from the reports are incorporated as a source of context both in preprocessing and in the evaluation and refinement processes of the models. The experiments were conducted with the GliNER model and evaluate that the use of metadata and pseudo-labeling can contribute to improving model performance on informal corpora, with positive impacts on automated information extraction in public security contexts. The results reinforce the potential of hybrid and domain-sensitive approaches for real-world Natural Language Processing applications in environments with scarce labeled data.

Student: Matheus dos Santos Moura

Title: Hybrid Anomaly and Change Point Detection for Pump-and-Dump Schemes in Centralized Cryptocurrency Exchanges

Advisors: Diogo Silveira Mendonça

Committee: Diogo Silveira Mendonça (Cefet/RJ), Eduardo Soares Ogasawara (Cefet/RJ) and Igor Machado Coelho (UFF)

Day/Hour: February 18, 2026 / 3 p.m.

Room: https://teams.microsoft.com/l/meetup-join/19%3ameeting_ZGFkNGM1MzMtOGMzNi00OWU5LTkzYjUtY2JhNGQxZmQzZjBl%40thread.v2/0?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%226821740b-ed93-4582-b3a3-b3bfbff6624e%22%7d

Abstract: The rapid growth of cryptocurrency markets has intensified concerns regarding market manipulation practices, particularly pump-and-dump schemes. Detecting such schemes remains challenging due to the high volatility of cryptocurrencies and the limited availability of reliable ground-truth data. Prior work has predominantly relied on anomaly detection techniques, which often exhibit limited precision and adaptability. In this work, we propose two offline statistical methods that explore a hybrid framework combining anomaly detection and change point detection for pump-and-dump detection. The first method, HD Pump, integrates volatility anomaly detection in price time series with change point detection applied to trading volume. The second method, HD Pump Plus, extends this approach by replacing the price time series with a rush-order-based time series. Experimental evaluation on a dataset of 178 confirmed pump-and-dump events from the Binance exchange shows that HD Pump Plus outperforms prior statistical approaches, achieving a precision of 96.4%, recall of 89.3%, and F1-score of 92.7%. These results demonstrate the effectiveness of hybrid detection strategies in advancing the state of the art while maintaining methodological simplicity.

Student: Edson Paulo da Silva Pinto Sobrinho

Title: Fine-tuning detection criteria to enhance anomaly identification in time series

Advisors: Eduardo Soares Ogasawara (advisor) and Kele Teixeira Belloze (co-advisor)

Committee: Eduardo Soares Ogasawara (CEFET/RJ), Kele Teixeira Belloze (CEFET/RJ), Rafaelli de Carvalho Coutinho (CEFET/RJ) and Esther Pacitti (INRIA / University of Montpellier)

Day/Hour: January 10, 2026 / 10:30 a.m.

Room: https://teams.microsoft.com/meet/24688731415374?p=YVHyDYwIW66rCysKq0

Abstract: Anomaly Detection (AD) is the problem of identifying observations that do not conform to typical ones in a time series. Detection methods implicitly define detection criteria, such as deviation measures, filter thresholds, and candidate anomaly selection strategies. Choosing inappropriate criteria results in inaccurate outputs, generating spurious alerts or missing events. Adjusting these criteria is essential for monitoring systems. To address this challenge, this study explores the fine-tuning of deviation measures, filter thresholds, and candidate selection strategies. Experimental results show that the proper choice of criteria significantly improves AD performance, often with greater impact than changing the detection methods.

Student: Luiz Cláudio Lemos de Oliveira

Title: Motif Detection in Time Series Using Autoencoders: An Analysis of Their Application to ECG Data

Advisor: Eduardo Soares Ogasawara

Committee: Eduardo Soares Ogasawara (CEFET/RJ), Laura Silva de Assis (CEFET/RJ), Helga Dolorico Balbi (CEFET/RJ) and Rebecca Pontes Salles (INRIA/FRA)

Day/Hour: January 08, 2026 / 10 a.m.

Room: https://teams.microsoft.com/meet/2899124353229?p=ykqgR5NeTJfPfXOhSF

Student: Michel Siqueira Reis

Title: Matching Detections to Events in Time Series with Computational Efficiency and Guaranteed Optimality

Advisors: Rafaelli Coutinho (Advisor) and Eduardo Ogasawara (Co-advisor)

Committee: Rafaelli Coutinho (Cefet/RJ), Eduardo Ogasawara (Cefet/RJ), Laura Assis (Cefet/RJ) and Rebecca Salles (INRIA)

Day/Hour: December 18, 2025 / 1 p.m.

Room: https://teams.microsoft.com/l/meetup-join/19%3ae20c8697654543fc9dd1e9924de5c2c0%40thread.tacv2/1763159776096?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%2254af42a0-5f30-4905-ac8d-10b96c6db26b%22%7d

Abstract: This work presents SmartSoftED, an optimized metric for evaluating the detection of point events in time series. The original metric, SoftED, introduces a “soft” evaluation based on temporal tolerance, assigning gradual scores to detections that occur near actual events. However, its current formulation relies on a greedy approach that does not guarantee optimality in all cases and incurs a quadratic computational cost, limiting its applicability in large-scale or real-time processing environments. SmartSoftED overcomes these limitations by introducing a strategy that decomposes the problem into manageable, disjoint subproblems: some can be solved efficiently without loss of optimality, while others are modeled as maximum-weighted matching problems on unbalanced bipartite graphs. This approach preserves the optimality of correspondences between detections and events while significantly reducing computational cost. In practice, the method achieves an average speedup of two orders of magnitude, making it suitable for certain large-scale applications and systems with strict temporal constraints.

Student: Fernando Henrique de Jesus Fraga da Silva

Title: Aprendizado por Reforço Profundo Aplicado à Negociação Intradiária de Múltiplas Ações

Advisors: Eduardo Bezerra da Silva (advisor) and Pedro Henrique González Silva (co-advisor)

Committee: Eduardo Bezerra da Silva (Cefet/RJ), Pedro Henrique González Silva (UFRJ), Aline Marins Paes Carvalho (UFF) e Glauco Fiorott Amorim (Cefet/RJ)

Day/Hour: December 8, 2025 / 3 p.m.

Room: https://teams.microsoft.com/v2/?meetingjoin=true#/l/meetup-join/19:PKOJTuK7mfHSDE6QkCWQCYp71f0xOMNoRgSUj4wjMKc1@thread.tacv2/1763760050816?context=%7b%22Tid%22%3a%228eeca404-a47d-4555-a2d4-0f3619041c9c%22%2c%22Oid%22%3a%22c03d6068-4733-48a6-bbb4-aa78f351d9cf%22%7d&anon=true&deeplinkId=91733be2-9804-4f09-ac6a-f1a362e67de8

Abstract: The stock market is a dynamic and volatile environment in which publicly traded companies negotiate fractions of their value, subject to continuous price fluctuations influenced by economic, political, and social factors. Anticipating these fluctuations is a complex task, especially in the context of intraday trading, where buy and sell decisions must be made within very short time intervals based on rapidly changing data. In this scenario, Reinforcement Learning (RL) emerges as a promising paradigm capable of developing adaptive strategies through the continuous interaction between agent and environment. This dissertation investigates the use of Deep Reinforcement Learning (DRL) techniques in financial trading, focusing on intraday scenarios involving multiple stocks. It proposes a DRL-based approach to estimate buy and sell actions simultaneously across various assets, using high-granularity market data to better approximate real trading conditions. Experimental analyses were conducted using the Proximal Policy Optimization (PPO) algorithm. The results indicate that the proposed agent outperformed traditional benchmark strategies, achieving gains exceeding 10 percentage points in certain cases.